Saliency Map

Sinlge number per feature in the form of a pixel-value. Often visualized as Heatmap to show which parts of an image are most important to classify it with a Neural Network.

A special type of Feature Attribution.

Model-Agnostic methods

• LIME

Gradient-based methods:

• Vanilla Gradient
• Grad-CAM

Path-attribution methods

• Deep Taylor
• Integrated Gradients
• LIME

Method that creates such an explanation is:

• Post-Hoc
• Saliency Map is result (subgroup of Feature Importance)
• Cant tell if Model-Agnostic or Model-Specific
• It is a Local Explanation

• Not Contrastive
  • Saliency Map shows where to look and not why
• It is Selectivity, most of the color is gray